Linear Algebra Refresher - web.eecs.umich.eduweb.eecs.umich.edu/~ocj/courses/autorob/autorob_05_linear_refresh.pdf · But, we need to start with a linear algebra refresher UM EECS

autorob.github.io

Linear Algebra Refresher

UM EECS 398/598 - autorob.github.io

ObjectiveGoal: Given the structure of a robot arm, compute – Forward kinematics: inferring the pose of the end-effector, given the state (angle) of each joint. – Inverse kinematics: inferring the joint states necessary to reach a desired end-effector pose.

But, we need to start with a linear algebra refresher


Reset: Kinematics

• State comprised of degrees-of-freedom (DOFs)

• DOFs describe translation and rotation axes of system

��

��

��

��

334 15. Physical Simulation

!"#$%&'"(!"#$%&'"( )*&+&('$+,-.'".#)*&+&('$+,-.'".#

! "#$%&''&()*+)%,-,()*.(+)/,0&123,4'

5 6,-,()*.(,&')4.00&43&()*+)7.,03'

5 892%243&%)$.'&)(&:,0&()*+)%.323,.0)

.:)7.,03'

5 ;&43.%)!<3=)%&$%&'&03')$.'&

23)2)-,>&0),0'3203).:)3,1&

!"#$%&'"(!"#$%&'"( )*&+&('$+,-.'".#)*&+&('$+,-.'".#

! "#$%&''&()*+)%,-,()*.(+)/,0&123,4'

5 6,-,()*.(,&')4.00&43&()*+)7.,03'

5 892%243&%)$.'&)(&:,0&()*+)%.323,.0)

.:)7.,03'

5 ;&43.%)!<3=)%&$%&'&03')$.'&

23)2)-,>&0),0'3203).:)3,1&Figure 15.11. Humanoid characters are composed of rigid bodies connected byjoints. Rotations about joints allow the character to pose in a variety of configurations(top). The character’s kinematics define degrees-of-freedom, specifying how jointscan rotate. Early research into physics-based animation by Hodgins et al. useddynamics that explicitly held the bones of humanoid character together (bottomleft). However, most dynamics engines used in modern games have independentrigid bodies that are held together by constraints, similar to how magnetism holdsGeomag structures together (bottom right).

a Lagrangian formulation24 or a Newton-Euler formulation25.Implicitly constrained dynamics, in contrast, does not explicitly enforce

constraints. Instead, constraints are expressed as implicit functions on thestate of the rigid bodies. Such constraints are satisfied when its functionevaluates to zero. The states where a constraint function evaluates tozero form an equilibrium region of state space. As states move furtherfrom this equilibrium region, their values, as evaluated by the constraintfunction, increase. Forces are generated from these constraint functions by

24The temporal derivative of the partial derivative of kinetic minus potential energywith respect to each generalized coordinate.

25At each time step, velocities and accelerations are propagated “forward” from theroot to outboard bodies and corrected “backward” to the root with forces and torquesbackward.


DOFs and Coordinate Spaces

• Each body has its own frame

• Joints connect two links (rigid bodies)

• e.g., hinge, prismatic, ball-socket

• A motor exerts force on a DOF axis

• Matrix transformations used to relate coordinate systems of bodies and joints

• Spatial geometry attached to each link, but does not affect the body’s coordinate frame

frames


hinge

prismatic

ball

Linear algebra

Notation


Planar 2-link Arm


SCARA Arm Selective Compliance Assembly Robot Arm

UM EECS 398/598 - autorob.github.iohttps://youtu.be/7X5Nmk85kQo

Motoman SK16

UM EECS 398/598 - autorob.github.iohttps://youtu.be/Wj17z5iSzEQ

Biped Hopper (MIT Leg Lab)

UM EECS 398/598 - autorob.github.iohttp://www.ai.mit.edu/projects/leglab/robots/robots.html

Big Dog (BDI)


Quad Rotor Helicopter

Safety is most importantUM EECS 398/598 - autorob.github.iohttps://youtu.be/0mDiH_ajStQ

ETH-Zurichhttps://www.youtube.com/watch?v=XxFZ-VStApo UM EECS 398/598 - autorob.github.io

PR2

How to express kinematics as the state of an articulated system? We need some math first.



What does algebra provide beyond arithmetic?

• Arithmetic applies to addition and multiplication of known numbers

• Algebra includes abstractions as variables

• Unknown numbers or expressions that can take on many values

• An algebra supports addition and multiplication of variables and numbers.

• For example, from: x2 = 5x – 6 • we get: (x – 2)(x – 3) = 0 • and thus: x = 2 or x = 3.


Many important complex systems are described by collections of linear equations. UM EECS 398/598 - autorob.github.io

What does is linear algebra provide beyond algebra?


• Describes spaces where vector operations are closed with respect to:

• addition

• scalar multiplication


• Describes spaces where vector operations are closed with respect to:

• addition

• scalar multiplication

Arithmetic Algebra Linear Algebra

Addition

Scalarmultiplication


Abstraction

Many important complex systems are described by collections of linear equations.

• Many important complex systems are described by collections of linear equations.

• An algebra of scalars, vectors, and matrices helps us work with these systems, keeping track of the complexity.

• Manipulate groups of known and unknown parameters, just like manipulating numbers.

• Linear algebra is essential for representing frames of reference, rotation, translation, and general 3D homogeneous transforms.


Linear Algebra (Rough) Breakdown• Geometry of Linear Algebra

• Vectors, matrices, basic operations, lines, planes, homogeneous coordinates, transformations

• Solving Linear Systems

• Gaussian Elimination, LU and Cholesky decomposition, over-determined systems, calculus and linear algebra, non-linear least squares, regression

• The Spectral Story

• Eigensystems, singular value decomposition, principle component analysis, spectral clustering

primary focus for AutoRob

needed for iterative IK


is solved by


=

is solved by

linear systems expressed in general matrix form as


=

vector of N unknowns to be found

M linear equations

each equation yields a hyperplane in N-D


=


M linear equations

If #unknowns > #equations, underdetermined system, usually with infinite solutions

If #unknowns < #equations, overdetermined system usually with no solutions

If #unknowns = #equations, usually has a unique solution



=


M linear equations

If #unknowns > #equations, underdetermined system, usually with infinite solutions

If #unknowns < #equations, overdetermined system, usually with no solutions

If #unknowns = #equations, usually has a unique solution



2D Exampleonly single point satisfies both lines


2D Example

any point on the line satisfies

only single point satisfies both lines no point satisfies

all three lines


3D Example

Each equation yields a 2D plane

in 3D space

A single point satisfies all equations


3D Example

How many solutions?


Coordinate Spaces (2D)• Two coordinate frames o0x0y0

and o1x1y1, and a point p.

• The location of point p can be described with respect to either coordinate frame: p0 = [5, 6]T and p1 = [-2.8, 4.2]T.

• The vector v1 is direction and magnitude from o0 to p, and v2 is from o1 to p.


Coordinate Spaces (2D)• Point p has a location.

• Vectors v1 and v2 have directions and magnitudes.

• v10 = [5, 6]T

• v11 = [7.77, 0.8]T

• v20 = [-5.1, 1]T

• v21 = [-2.8, 4.2]T Note: Vectors can only be added when they are in the same coordinate frame.

vector 1 in frame 0

vector 1 in frame 1

vector 2 in frame 0

vector 2 in frame 1


Coordinate Spaces (2D)• Point p has a location.

• Vectors v1 and v2 have directions and magnitudes.

• v10 = [5, 6]T

• v11 = [7.77, 0.8]T

• v20 = [-5.1, 1]T

• v21 = [-2.8, 4.2]T Note: Vectors can only be added when they are in the same coordinate frame.UM EECS 398/598 - autorob.github.io

Vectors and Matrices

N-dimensional vector M-by-N matrix


2D Vector

var a = [ [2], [3] ];

A vector is a motion in space


3D Vector

var a = [ [ax], [ay], [az] ];


Vector Addition and Subtraction

vector addition is order independent

vector addition visually

vector subtraction is additionwith negated vector

vector result


Magnitude and Unit VectorThe magnitude of a vector is the square root of the sum of

squares of its components

A unit vector has a magnitude of one.Normalization scales a vector to unit length.

A vector can be multiplied by a scalar


Dot Product

Measures the similarity in direction of two vectors

scalar result


Projections

Scalar projection of one vector onto another

Vector projection

Dot products related to projections onto vectors.


is unit length

Checkpoint

• What is the dot product of a vector with itself?

• the square of the vector length

• What is the dot product of two orthogonal vectors?

• 0


Checkpoint

• What is the dot product of a vector with itself?

• the square of the vector magnitude

• What is the dot product of two orthogonal vectors?

• 0


Checkpoint

• How many unit vectors are perpendicular to a 2D vector?

• 2 (positive and negative)


• Infinite and lie in plane


Checkpoint


• 2 (positive and negative)


• Infinite and lie in plane


Given two vectors, how to compute a vector orthogonal to both?


Cross Product

Results in new vector c orthogonal toboth original vectors a and b

Length of vector c is equal to area of parallelogram formed by a and b

Assumes a and b are in same frame


Relating linear and angular velocity

angular velocity for rotation of p about vector k

vector to a point p

linear velocity of point p

r

v

k

p


Matrices• A Matrix is a rectangular array of numbers

var mat = [ [1, 0, 0, 0], [0, 1, 0, 0], [0, 0, 1, 0],

[0, 0, 0, 1] ];

What is this matrix?


Matrix-vector multiplication


For example


For example


Matrix-vector multiplication (two interpretations)

1) Row story: dot product of each matrix row

2) Column story: linear combination of matrix columns


Revisiting the cross product: Skew-symmetric matrices

A given 3D vector

can be expressed as a skew-symmetric matrix

such that the cross product with another vector is a matrix multiplication


Linear SystemsWe can use a variable instead of a vector, which gives us a linear system.

Enabling the general form:


Matrices• A Matrix is a rectangular array of numbers

var mat = [ [1, 0, 0, 0], [0, 1, 0, 0], [0, 0, 1, 0],

[0, 0, 0, 1] ];

var mat = [ [1, 0, 0, tx], [0, 1, 0, ty], [0, 0, 1, tz],

[0, 0, 0, 1] ];

What is this matrix?


Translation matrix example

jsmat for Assignment 3


Translation matrix example

jsmat for Assignment 3


Matrix Geometry: Column Story


Matrix Multiplication• Scalar Multiplication

• Multiplication of two matrices

Each entry of product matrix AB is a dot product of a row of A with a column of B


Matrix multiplication

Do this dot product for each row/column combination

AB

AB

row 3 of Acol 4 of B

Finger sweeping rule should be second nature! - Left finger sweeps left to right - Right finger sweeps top to bottom


Matrix Multiplication Reminders

• Number of columns of A must match number of rows of B

• Multiplying a (MxK) matrix with a (KxN) matrix will produce an (MxN) matrix

• Matrix multiplication is not commutative: AB != BA


Example

(1, 2, 3) • (7, 9, 11) = 1×7 + 2×9 + 3×11 = 58

(1, 2, 3) • (8, 10, 12) = 1×8 + 2×10 + 3×12 = 64

(4, 5, 6) • (7, 9, 11) = 4×7 + 5×9 + 6×11 = 139

(4, 5, 6) • (8, 10, 12) = 4×8 + 5×10 + 6×12 = 154


For example


For example


Checkpoint

• Which of the following matrix multiplications are valid?



Notable Matrices and Operations• Matrix identity (I) causes no change: A = ImA = AIn

• Diagonal elements Aii = 1

• Off-diagonal elements Aij = 0, i≠j

• Matrix inverse (A-1): if AA-1 = A-1A = I

• Distributing matrix inverse: (AB)-1 = B-1A-1

• Matrix transpose (AT): a matrix’s reflection about its diagonal

• Distributing matrix transpose: (AB)T = BTAT


Matrix Geometry: Row Story


Solving linear systems

What would be the direct way to solve for x?

Invert A an multiply by b

Can this always be done?

Can we approximate the solution?

Pseudoinverse least-squares approximation




Invert A and multiply by b


Matrix rank and inversion

• Let A be a square n by n matrix. A is invertible if full rank and a matrix B exists such that

• Rank of a matrix A is the size of the largest collection of linearly independent columns of A

• A is invertible (nonsingular) if it has full rank

• Gaussian elimination can find matrix inverse

• Singular matrix cannot be inverted this way


Solution by Decomposition• In real applications, inverse not computed to solve linear systems

• Efficiency, numerical precision, etc.

• Matrix decomposed into product of lower and upper triangular matrices

• LU decomposition

• Cholesky decomposition

• Permits finding solution by forward substitution followed by backward substitution





Can we approximate the solution?







No. But, we can approximate. How?




Pseudoinverse

• For matrix A with dimensions N x M with full rank

• Find solution that minimizes squared error:

• Left pseudoinverse, for when N > M, (i.e., “tall”)

• Right pseudoinverse, for when N < M, (i.e., “wide”)

s.t.

s.t.


Polynomial Regression• Given n data points as input-output (xi,yi), estimate parameters β of

best fitting m-order polynomial:

• Model in matrix form:

• each data point forms a row

• Solve for least squares best fit:


Julien’s Demohttps://herojeanpierre.github.io/Least_Squared_Best_Fit/

Forward Kinematics

autorob.github.io

Next Class


Linear Algebra Refresher - web.eecs.umich.eduweb.eecs.umich.edu/~ocj/courses/autorob/autorob_05_linear_refresh.pdf · But, we need to start with a linear algebra refresher UM EECS

Documents